A power distribution network intelligent inspection and fault diagnosis method based on multi-source information fusion

Abstract

The application belongs to the technical field of smart grid fault detection and identification, and relates to a power distribution network intelligent inspection and fault diagnosis method based on multi-source information fusion. The method comprises an unmanned aerial vehicle inspection platform, an airborne edge computing terminal, a cloud operation and management platform, a PMS system and a mobile terminal. The application synchronously collects visible light images, infrared thermal images and independent temperature sensing data through the unmanned aerial vehicle platform, and utilizes deep learning and multi-modal fusion technology to realize data complementation and collaborative analysis at a feature level, so that real-time and accurate identification of power distribution network lines and equipment faults is finally completed.

Description

Technical Field

[0001] This invention belongs to the field of smart grid fault detection and identification technology, and relates to a method for intelligent inspection and fault diagnosis of distribution networks based on multi-source information fusion. Background Technology

[0002] As the "last mile" of power transmission, the safe and stable operation of the power distribution network is crucial. Traditional inspections rely on manual visual inspection or single sensors, resulting in low efficiency, high risk, and a high risk of missed detections. With the development of drones and artificial intelligence technologies, vision-based automated inspections have become the mainstream direction. However, current technologies still face severe challenges: First, recognition capabilities are limited by the environment. Visible light cameras fail at night, in rain, fog, or strong backlight; while infrared thermal imagers can detect heat, they cannot identify non-thermal defects such as mechanical damage. Second, multi-source information is not deeply integrated. Existing methods mostly use simple data overlay or post-decision fusion, failing to achieve feature-level complementarity and enhancement, resulting in low information utilization. Third, there is a lack of deep coupling with physical states (such as temperature). Temperature is a key indicator for judging electrical faults, but most current vision systems only use it as an auxiliary criterion, without spatial mapping and collaborative optimization with image features. Fourth, the system's real-time performance and integration are insufficient. Complex algorithms often rely on the cloud, making it difficult to meet the real-time processing requirements of airborne edge devices, and data is not shared with existing power grid business systems, forming "information silos." Therefore, there is an urgent need for a new fault identification method that can deeply integrate multimodal sensing data, adapt to complex environments, achieve edge intelligence, and efficiently link with the power grid system. Summary of the Invention

[0003] This invention addresses the problems existing in the traditional multimodal information fusion process by proposing a method for intelligent inspection and fault diagnosis of power distribution networks based on multi-source information fusion.

[0004] This invention presents a technical solution for automatic fault identification by simultaneously acquiring visible light images, infrared thermal images, and temperature data using a drone platform, and utilizing cross-modal feature fusion and temperature collaborative analysis techniques. This method integrates computer vision, multimodal deep learning, sensor fusion, and edge computing technologies, and is applicable to real-time fault detection, hazard warning, and status assessment of distribution network lines, towers, insulators, and auxiliary equipment. It can be widely applied in scenarios such as daily inspections, emergency fault response, and intelligent operation and maintenance management of power systems, effectively improving the safety and reliability of power grid operation. Its core lies in simultaneously acquiring visible light images, infrared thermal images, and independent temperature sensor data using a drone platform, and utilizing deep learning and multimodal fusion technologies to achieve data complementarity and collaborative analysis at the feature level, ultimately achieving real-time and accurate identification of faults in distribution network lines and equipment.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution: A distribution network intelligent inspection and fault diagnosis system based on multi-source information fusion, the system includes a drone inspection platform, an airborne edge computing terminal, a cloud-based operation and maintenance management platform, a PMS system, and a mobile terminal.

[0006] The UAV inspection platform includes a visible light camera for identifying non-thermal defects, an infrared thermal imager for detecting overheating faults, and an independent temperature sensor for real-time acquisition of multi-point time-series temperatures of the environment and equipment surfaces. The airborne edge computing terminal includes a data preprocessing module for standardizing, enhancing, and aligning multi-source data; a cross-modal feature fusion module for extracting and fusing deep features from visible light and infrared images; a temperature collaborative analysis module that combines time-series temperature data and infrared thermal images to generate a temperature anomaly heatmap; a fault identification and classification module that uses an improved lightweight YOLOv8 network to achieve end-to-end fault detection and classification; and an encrypted communication module that encrypts and uploads the identification results to the cloud. The cloud-based operation and maintenance management platform is linked with the PMS system to automatically generate maintenance work orders and push early warning information, inspection reports, and alarms to the fault identification and classification module and the mobile terminal, respectively.

[0007] The method for intelligent inspection and fault diagnosis of power distribution network using the above system for multi-source information fusion is as follows: (1) Enhance and normalize the image acquired by the visible light camera. For the visible light image, the adaptive histogram equalization and multi-scale Retinex algorithm are used to enhance the weak light and strong reflective areas. The formula is used to enhance the weak light and strong reflective areas. The enhanced visible light image is obtained, where G k For Gaussian kernel, W k Here, x and y represent the horizontal and vertical coordinates of the pixel, respectively. This represents the pixel value at the same coordinates (x, y) in the image after enhancement processing. This represents the pixel value of the original image at coordinates (x, y).

[0008] (2) Using ResNet-18 to extract multi-scale features from the enhanced visible light image to obtain Infrared branch features are extracted from infrared images using the lightweight MobileNetV3 algorithm, and then processed through a feature alignment network to refine the infrared features. Perform a spatial transformation to output an infrared feature map aligned with the visible light feature space. (Aligned using this formula:) ,in, For transformation function, To transform parameters (through prediction via a learnable network), this process ensures that the features of each modality strictly correspond at the pixel level during subsequent fusion, which is a key step in the deep fusion of cross-modal features.

[0009] (3) After that and The data is fed into the cross-modal attention fusion module, where the feature weights for each modality are dynamically calculated. The weights for visible light and infrared features are calculated as follows: ,in This represents the weights assigned to visible light features. This represents the weight assigned to the infrared feature. It is a learnable matrix.

[0010] To flexibly respond to different scenarios, dynamic weighted fusion is adopted. and The calculation formula is as follows: Based on this feature, the fused features are obtained. .

[0011] (3) To achieve deep fusion of temperature data and visual features, a sliding window analysis is performed on the time-series temperature data collected by independent sensors to calculate the temperature gradient and statistical anomalies: ,in This represents the average of all temperature data within a sliding time window preceding the current moment, indicating a recent normal temperature level. It is the absolute deviation of the current temperature value from the recent average level. If... If it is, then it is marked as a temperature anomaly point. It is a threshold that is preset specifically for temperature anomaly detection tasks.

[0012] (4) When At that time, the infrared image is converted into a temperature matrix. Generate a thermal map of temperature anomalies. : ,in For ambient temperature, This is a temperature matrix. This is the highest temperature currently seen in infrared images. This is the lowest temperature currently seen in infrared images. This is a sigmoid function. The final result is a heatmap of temperature anomalies, the same size as the infrared image. The brightness value (between 0 and 1) of each pixel in the image does not represent absolute temperature, but rather the probability of abnormal heating at that location. Brighter areas have a higher probability of abnormal heating.

[0013] (5) Visual fusion features are generated through the cross-modal feature fusion module, spatial mapping and fusion are performed, and the temperature anomaly heat map and visual fusion features are spliced ​​together in the channel dimension and input into the subsequent recognition network.

[0014] (6) Finally, spatial mapping and fusion are performed to convert the heat map into a single image. Visual fusion features Channel splicing is performed to obtain The data is then input into the subsequent recognition network.

[0015] (7) The recognition algorithm uses an improved lightweight YOLOv8 network. This module deeply integrates the temperature information obtained from infrared thermometry as a supervisory and guiding signal into the training and inference process of the target detection network, constructing an intelligent detection system with "visual-thermal imaging" collaborative analysis. This significantly improves the reliability and anti-interference ability of the recognition of heat-generating defects, especially. (8) For the features after channel splicing It is necessary to first upsample it to a uniform size before subsequent calculations can be performed. This operation yields , This indicates upsampling. To generate features that are both rich in strong semantics and high-resolution details, providing high-quality input for subsequent detection heads, an upsampling method is employed. By splicing lateral connectivity feature maps from the same scale as the backbone network, more detailed information can be preserved. This represents the convolution operation. For splicing, C lateral This represents a feature map from the same scale as the backbone network.

[0016] (9) After that The data is fed into the detection head, which is responsible for making the final prediction. This model extends this prediction into three parallel branches, achieving a "triple decision" of classification, localization, and temperature verification. The classification branch predicts the fault category of the target within the bounding box, outputting a probability distribution. The bounding box regression branch accurately predicts the position and size (center point coordinates, width, and height) of the target bounding box. For each predicted bounding box, the temperature verification branch simultaneously predicts a local temperature rise value. This value represents the amount of temperature increase the model "thinks" the area will experience relative to normal. During training and inference, the system obtains the measured temperature rise value of the area from the infrared sensor. Consistency verification is performed by comparing the predicted local temperature rise value with the measured temperature rise value.

[0017] (10) The loss function is the "command stick" for model training; multi-task models require careful balancing of the weights of each sub-task. The formula for the total loss function is: ,in The classification loss measures the accuracy of the category prediction, ensuring that the model can correctly identify the type of fault. It is the bounding box regression loss, which measures the deviation between the predicted box position and the true box position. It ensures that the model can accurately locate the fault. This represents the temperature regression loss. and These are hyperparameter weighting coefficients. The importance of controlling positioning accuracy in the overall objective. This emphasizes the importance of temperature prediction consistency in the overall objective. By adjusting... This determines the extent to which the model can "trust" and fit the temperature signal.

[0018] Ultimately, the network output is no longer simply "There is a suspected overheating fault here," but rather "There is a target here, its visual morphology is classified as joint overheating, and both prediction and actual measurement show that its temperature rise reaches X degrees, therefore it is confirmed as a high-reliability alarm." Therefore, the multimodal information fusion proposed in this invention has the core value characteristics of precision and interpretability in industrial defect detection.

[0019] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention uses a drone platform to simultaneously collect visible light images, infrared thermal images, and independent temperature sensor data. By utilizing deep learning and multimodal fusion technology, it achieves data complementarity and collaborative analysis at the feature level, ultimately enabling real-time and accurate identification of faults in power distribution network lines and equipment. Detailed Implementation

[0020] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described below with reference to specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0021] Numerous specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways than those described herein, and therefore the invention is not limited to the specific embodiments disclosed in the following specification. Example

[0022] The system provided in this embodiment consists of three layers: a UAV inspection platform, an airborne edge computing terminal, and a cloud-based operation and management platform. The UAV inspection platform is equipped with a visible light camera to acquire high-resolution RGB images for identifying non-thermal defects such as external damage, dirt, and bird nests; an infrared thermal imager acquires temperature distribution images for detecting thermal faults (such as overheated connectors and partial discharge); independent temperature sensors collect real-time temperatures at multiple points on the environment and equipment surfaces, providing high-precision, low-noise time-series temperature data; and a synchronization trigger unit ensures strict spatiotemporal synchronization of multi-sensor data through hardware signals. The airborne edge computing terminal integrates the following modules: a data preprocessing module standardizes, enhances, and aligns multi-source data; a cross-modal feature fusion module extracts and fuses deep features from visible light and infrared images; a temperature collaborative analysis module combines time-series temperature data and infrared thermal images to generate a temperature anomaly heatmap; and a fault identification and classification module, based on an improved lightweight YOLOv8 network, achieves end-to-end fault detection and classification. The identification results are then encrypted and uploaded to the cloud via an encrypted communication module. The cloud-based operation and maintenance management platform works in conjunction with the PMS system to automatically generate maintenance work orders and push early warning information, inspection reports and alarms to the fault monitoring system and mobile terminals respectively.

[0023] To ensure the effectiveness of multi-source data fusion, the system first performs data preprocessing and spatiotemporal alignment. Image enhancement and normalization: For visible light images, adaptive histogram equalization (CLAHE) and multi-scale Retinex algorithm are used to enhance weak light and highly reflective areas. .

[0024] Among them G k For Gaussian kernel, W k As weights, the infrared image undergoes non-uniformity correction and temperature calibration, and is converted into a temperature matrix T. ir (x, y).

[0025] The synchronous triggering unit ensures that the timestamps of the image and temperature data are consistent for spatiotemporal synchronization. Based on feature point matching and thin plate spline transformation (TPS), pixel-level spatial alignment of visible light and infrared images is achieved, which is more accurate than traditional affine transformation.

[0026] The cross-modal feature fusion module employs a dual-branch feature extraction + attention fusion mechanism. The visible light branch uses ResNet-18 to extract multi-scale features. The infrared branch uses the lightweight MobileNetV3 to extract feature F. ir This approach balances accuracy and speed. Then, a Spatial Transformer Network (STN) is used to spatially transform the infrared features, aligning them semantically with the visible light features. The calculation formula is as follows: .

[0027] Where r is the transformation function and θ is the transformation parameter, which is predicted by a learnable network.

[0028] After that and The data is fed into the cross-modal attention fusion module, where the feature weights for each modality are dynamically calculated. The weights for visible light and infrared features are calculated as follows: ,in This represents the weights assigned to visible light features. This represents the weight assigned to the infrared feature. It is a learnable matrix.

[0029] To flexibly respond to different scenarios, dynamic weighted fusion is adopted. and The calculation formula is as follows: Based on this feature, the fused features are obtained. .

[0030] This dynamic weighting mechanism allows the model to flexibly adapt to different scenarios. In well-lit scenes, visible light features are given higher weights; in low-light or hot-target scenes, infrared features are relied upon more. Compared to simple averaging or fixed-weight fusion, this attention-driven fusion method can improve the robustness and accuracy of the model under different conditions.

[0031] The temperature co-analysis module enables deep fusion of temperature data and visual features. It integrates time-series temperature data {T} collected from independent sensors. t Perform sliding window analysis to calculate the temperature gradient and statistical anomalies. The calculation formula is as follows: Where μ window This represents the average of all temperature data within a sliding time window preceding the current moment, indicating a recent normal temperature level, |Tt-μ window | represents the absolute deviation of the current temperature from the recent average level. The larger this value, the further the current temperature deviates from "normal." σ window This represents the standard deviation of temperature data within the same sliding window. It measures the range of recent temperature fluctuations. A large value indicates significant temperature volatility, while a small value indicates stable temperature. If ΔT > 0. If it is, then it is marked as a temperature anomaly point. It is a threshold that is preset specifically for temperature anomaly detection tasks.

[0032] when At that time, the infrared image is converted into a temperature matrix. Generate a thermal map of temperature anomalies. : ,in For ambient temperature, This is a temperature matrix. This is the highest temperature currently seen in infrared images. This is the lowest temperature currently seen in infrared images. This is a sigmoid function. The final result is a heatmap of temperature anomalies, the same size as the infrared image. The brightness value (between 0 and 1) of each pixel in the image does not represent absolute temperature, but rather the probability of abnormal heating at that location. Brighter areas have a higher probability of abnormal heating.

[0033] Visual fusion features are generated through a cross-modal feature fusion module, and then spatially mapped and fused. The temperature anomaly heatmap and visual fusion features are concatenated along the channel dimension and input into the subsequent recognition network. Finally, spatial mapping and fusion are performed to transform the heatmap H... temp Visual fusion feature F fused Channel splicing is performed to obtain The data is then input into the subsequent recognition network.

[0034] In power equipment inspection, traditional visual inspection algorithms (YOLOv8) primarily rely on visible light features such as shape, color, and texture. However, many critical defects (such as overheated joints and localized insulator leakage) may not be morphologically obvious, but their abnormal temperature is a key characteristic. The core innovation of this module lies in deeply integrating temperature information obtained from infrared thermography as a supervisory and guiding signal into the training and inference process of the target detection network, constructing an intelligent detection system that combines visual and thermal imaging analysis. This significantly improves the reliability and anti-interference capability of identifying, especially, heat-generating defects.

[0035] Features after channel splicing It is necessary to first upsample it to a uniform size before subsequent calculations can be performed. This operation yields , This indicates upsampling. To generate features that are both rich in strong semantics and high-resolution details, providing high-quality input for subsequent detection heads, an upsampling method is employed. This operation, which stitches together lateral connectivity feature maps from the same scale as the backbone network, can retain more detailed information. This represents the convolution operation. For splicing, C lateral This represents a feature map from the same scale as the backbone network.

[0036] Where P up The upsampled feature map from the previous layer has a higher resolution, but may be relatively coarse. (C) lateralThe lateral connectivity feature map, derived at the same scale as the backbone network, retains more detailed information. This operation generates features at each layer of the feature pyramid that are both rich in strong semantics and high-resolution details, providing high-quality input for subsequent detection heads.

[0037] This is an improvement on standard upsampling operations. In power scenarios, temperature anomaly areas (hot spots) are often critical small targets. The model dynamically adjusts the attention during the upsampling process based on the distribution map of temperature anomaly areas (obtained from infrared image processing). Simply put, the network "learns" to perform more refined feature reconstruction in areas with significantly higher temperatures, thereby increasing the resolution and sensitivity of features in these areas. This allows even minor overheating defects to be clearly represented on the feature map, avoiding missed detections.

[0038] After that The data is fed into the detection head, which is responsible for making the final prediction. This model extends this prediction into three parallel branches, achieving a "triple decision" of classification, localization, and temperature verification. The classification branch predicts the fault category of the target within the bounding box, outputting a probability distribution. The bounding box regression branch accurately predicts the position and size (center point coordinates, width, and height) of the target bounding box. For each predicted bounding box, the temperature verification branch simultaneously predicts a local temperature rise value. This value represents the amount of temperature increase the model "thinks" the area will experience relative to normal. During training and inference, the system obtains the measured temperature rise value of the area from the infrared sensor. Consistency verification is performed by comparing the predicted local temperature rise value with the measured temperature rise value.

[0039] The loss function acts as a "command stick" for model training; multi-task models require careful balancing of the weights of each sub-task. The formula for the total loss function is: . Where L cls The classification loss measures the accuracy of category predictions, ensuring the model correctly identifies the fault type. box It is the bounding box regression loss, used to measure the deviation between the predicted box position and the ground truth box position. It ensures that the model can accurately locate faults. temp λ represents the temperature regression loss. box and λ temp These are the hyperparameter weighting coefficients, λ box The importance of control positioning accuracy in the overall objective, λ temp This emphasizes the importance of temperature prediction consistency in the overall objective. This can be achieved by adjusting λ. temp This determines the extent to which the model should "trust" and fit the temperature signal.

[0040] Ultimately, the network output is no longer just "There is a suspected overheating fault here," but rather "There is a target here, its visual morphology is classified as joint overheating, and both prediction and actual measurement show that its temperature rise reaches X degrees, therefore it is confirmed as a high-reliability alarm." This precisely reflects the core value of multimodal information fusion in industrial defect detection, moving towards greater precision and interpretability.

[0041] The hardware deployment of this invention constitutes a complete three-layer system architecture. At the front end is the UAV inspection platform, employing a heavy-load, long-endurance industrial-grade multi-rotor UAV. Its gimbal integrates core sensing units: a high-resolution zoom visible light camera for capturing shape details, a high-precision infrared thermal imager for acquiring temperature distribution images, and a set of high-response independent digital temperature sensors for direct contact or close-range measurement of key equipment temperatures. To ensure data consistency, all sensors are linked at the hardware level through a synchronization trigger unit based on a high-precision timing signal, achieving millisecond-level timestamp synchronization. The middle layer is a ruggedized airborne edge computing terminal mounted on the UAV fuselage. Its core is an embedded device with powerful AI computing capabilities, pre-installed with all algorithm software, responsible for real-time task processing. The back end is a cloud-based operation and maintenance management platform deployed in the power company's data center. This platform interconnects with existing power distribution network production management systems, fault monitoring systems, and mobile applications through standardized interfaces, responsible for aggregating information, scheduling operations, and forming an operation and maintenance closed loop.

[0042] In actual operation, the system first performs automated task planning and execution. Maintenance personnel, based on the power grid geographic information system and equipment ledgers, plan the drone's refined inspection routes and key inspection targets on the cloud platform. The drone autonomously flies to the predetermined location, such as above a power pole or insulator. At this point, the synchronization trigger unit is activated, directing visible light cameras, infrared thermal imagers, and temperature sensor arrays to perform strictly synchronized data acquisition on the same target. The acquired multi-source raw data is then transmitted to the onboard edge computing terminal via a high-speed link.

[0043] Data immediately enters the real-time processing pipeline at the edge terminal. The preprocessing stage adaptively enhances the visible light image to address lighting challenges, calibrates the infrared image, and utilizes advanced spatial transformation algorithms to achieve pixel-level precise alignment between the visible light and infrared images, while simultaneously associating temporal temperature flow with image frames. Next, the aligned dual-light image is fed into a cross-modal feature fusion module, where features are extracted using a dual-branch deep network, and an attention mechanism is used to dynamically weigh and fuse visible light and infrared information to generate a robust joint visual representation of environmental changes.

[0044] Simultaneously, the temperature co-analysis module operates independently, performing real-time sliding window statistics on continuous temperature data from the sensors to intelligently identify abnormal heat points and generate a high-brightness temperature anomaly probability heatmap by combining it with an infrared temperature matrix. This heatmap, along with the aforementioned fused visual features, is input into an improved lightweight target detection network. This network uniquely integrates a temperature-guided mechanism and a triple-decision detection head, ultimately outputting reliable diagnostic results including fault type, precise location, and quantified temperature rise.

[0045] The processed structured diagnostic results are encrypted and transmitted back to the cloud platform in real time. The platform automatically associates fault information with equipment assets, generates maintenance work orders with one click, and pushes them to the mobile terminals of maintenance personnel. At the same time, it sends detailed early warning reports to the monitoring system and management. After maintenance personnel complete on-site handling according to the work orders, they provide feedback on the results via their mobile terminals. The cloud platform then updates the equipment status, thus completing a fully automated closed loop from intelligent sensing and accurate diagnosis to maintenance response and feedback, greatly improving the efficiency and intelligence level of power grid operation and maintenance.

[0046] This embodiment achieves multi-level, adaptive cross-modal feature deep fusion, significantly improving recognition accuracy and robustness. The invention realizes the interaction and alignment of visible light, infrared, and temperature data at the feature level of the deep neural network. Through a cross-modal attention mechanism, it dynamically adjusts the fusion weights, enabling the model to adaptively select the dominant modality based on conditions such as lighting and scene. This overcomes the information loss problem caused by simple superposition of multimodal information or decision-level fusion in existing methods, significantly improving the detection accuracy and environmental adaptability for non-heating mechanical defects and heating electrical defects. It enhances the system's recognition stability in complex lighting and harsh environments. By introducing adaptive image enhancement (such as CLAHE and multi-scale Retinex) and lighting perception fusion strategies, the system can maintain reliable recognition capabilities under extreme conditions such as extremely low light, strong backlight, rain, and fog, solving the problem of traditional visible light or single infrared modal failure in specific environments and significantly improving the all-weather operation capability of the inspection system. This system achieves deep coupling of temperature information and visual features, improving the reliability of detecting heat-generating defects. By integrating high-precision independent temperature sensors and combining time-series temperature anomaly detection with spatial thermal map generation, temperature information is integrated as a guiding signal into the target detection network, forming a "vision-temperature" collaborative verification mechanism. This enables the system not only to detect heat-generating faults but also to perform quantitative analysis and credibility assessment of abnormal temperature rises, improving the interpretability and reliability of alarms. It possesses real-time edge processing capabilities, supporting low-latency response for UAV inspections. Through lightweight network design (such as MobileNetV3 and improved YOLOv8), the model significantly reduces computational complexity while maintaining accuracy, enabling end-to-end real-time fault identification and classification on airborne edge computing terminals. This meets the high real-time requirements of UAV inspections and overcomes the high latency and slow response issues caused by existing methods relying on cloud computing. A systematic and closed-loop intelligent operation and maintenance system has been constructed, enabling efficient linkage with the power grid business system. Through standardized data interfaces and communication protocols, the system can automatically push the identification results to the cloud operation and maintenance platform and connect with the distribution network production management system (PMS), fault monitoring system, etc., supporting automated work order generation, early warning information push and decision support. It realizes a closed loop of the whole process from fault identification to maintenance response, effectively solves the "information silo" problem, and improves the overall efficiency and intelligence level of power grid operation and maintenance.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments for application in other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A power distribution network intelligent inspection and fault diagnosis system based on multi-source information fusion, characterized in that, The system includes a drone inspection platform, an airborne edge computing terminal, a cloud-based operation and maintenance management platform, a PMS system, and a mobile terminal. The drone inspection platform includes a visible light camera for identifying non-thermal defects, an infrared thermal imager for detecting overheating faults, and independent temperature sensors for real-time acquisition of multi-point time-series temperatures of the environment and equipment surfaces. The airborne edge computing terminal includes a data preprocessing module for standardizing, enhancing, and aligning multi-source data; a cross-modal feature fusion module for extracting and fusing deep features from visible light and infrared images; a temperature collaborative analysis module for combining time-series temperature data and infrared thermal images to generate a temperature anomaly heatmap; and a fault identification and classification module for end-to-end fault detection and classification based on an improved lightweight YOLOv8 network. An encrypted communication module that encrypts and uploads the recognition results to the cloud; The cloud-based operation and maintenance management platform works in conjunction with the PMS system to automatically generate maintenance work orders and push early warning information, inspection reports and alarms to the fault identification and classification module and mobile terminals, respectively.

2. A method for intelligent inspection and fault diagnosis of a distribution network using the system described in claim 1 for multi-source information fusion, characterized in that, The steps are as follows: (1) Enhance and normalize the image acquired by the visible light camera. Adaptive histogram equalization and multi-scale Retinex algorithm are used to enhance the weak light and strong reflective areas to obtain the enhanced visible light image. (2) Using ResNet-18 to extract multi-scale features from the enhanced visible light image to obtain Infrared branch features are extracted from infrared images obtained by infrared thermal imagers using the lightweight MobileNetV3 algorithm. Then, a feature alignment network is used to spatially transform the infrared branch features, outputting an infrared feature map that is spatially aligned with visible light features. ; (3) After that and The data is fed into the cross-modal attention fusion module, where the feature weights for each modality are dynamically calculated. The calculation formula is ,in This represents the weights assigned to visible light features. This represents the weight assigned to the infrared feature. It is a learnable matrix; and For dynamic weighted fusion, further formulas are used. The resulting visual fusion features were obtained. ; (3) Perform sliding window analysis on the time-series temperature data collected by independent temperature sensors, calculate the temperature gradient and statistical anomalies, and achieve deep fusion of temperature data and visual features: The formula is as follows: ,in This represents the average value of all temperature data within a sliding time window preceding the current moment. It is the absolute deviation of the current temperature value from the recent average level; if If so, mark it as a temperature anomaly point; The preset threshold for the temperature anomaly detection task; (4) When At that time, the infrared image is converted into a temperature matrix. Generate a thermal map of temperature anomalies. T env For ambient temperature, T ir (x,y) is the temperature matrix, T max This is the highest temperature currently seen in infrared images, T min This is the lowest temperature currently seen in infrared images. The Sigmoid function is used to obtain a temperature anomaly probability heatmap of the same size as the infrared image. The brightness value of each pixel in the image represents the probability of abnormal heating at that location. The brighter the area, the higher the probability of abnormal heating. (5) Visual fusion features are generated through the cross-modal feature fusion module, and spatial mapping and fusion are performed to combine the temperature anomaly probability heatmap with the visual fusion features F. fused Perform channel-dimensional splicing and input it into the subsequent recognition network; (6) Perform spatial mapping and fusion to transform the temperature anomaly heatmap H temp (x,y) and visual fusion features F fused F is obtained by splicing the channel dimensions. joint The data is then input into the subsequent recognition network. (7) By using an improved lightweight YOLOv8 network recognition algorithm, the temperature information obtained by infrared thermometry is used as a supervision signal and a guidance signal and integrated into the training and inference process of the target detection network to build an intelligent detection system with visual-thermal imaging collaborative analysis. (8) For the feature F after channel splicing joint Samples are taken to a uniform size for subsequent calculations; the calculation formula is as follows: ; Indicates upsampling; using The concatenation is performed using lateral connectivity feature maps of the same scale as the backbone network, where Conv represents the convolution operation, Concat represents concatenation, and C... lateral This represents a feature map from the same scale as the backbone network; (9) P out The data is fed into the detection head, where a final prediction is made and a probability distribution is output. (10) The accuracy of category prediction is measured using the total loss function formula: L cls For classification loss, L box It is the bounding box regression loss, L temp For temperature regression loss; and These are hyperparameter weighting coefficients; adjusted by... This increases the reliability of the model.

3. The method for intelligent inspection and fault diagnosis of distribution networks based on multi-source information fusion according to claim 2, characterized in that, In step (1), adaptive histogram equalization and multi-scale Retinex algorithm are used to enhance weak light and strong reflective areas. The calculation formula for the enhanced visible light image is as follows: Among them G k For Gaussian kernel, W k Here, x and y represent the horizontal and vertical coordinates of the pixel, respectively. This represents the pixel value at the same coordinates (x, y) in the image after enhancement processing. This represents the pixel value of the original image at coordinates (x, y).

Images